[ntp:questions] .1 Microsecond Synchronization
david at ex.djwhome.demon.co.uk.invalid
Thu Jun 4 21:07:17 UTC 2009
> The company I am working for needs to be able to record timestamps in a trading
> system logs down to a .1 microsecond accuracy.
I suppose they want this on Windows as well. It's definitely not
possible on Windows. 100ns resolution may be achievable, but not using
standard Windows clock sources, and it would be pointless, as the
scheduling uncertainty will probably be in the 10s of milliseconds.
> We will have servers located in London, New York and Chicago. There will be a
> dedicated resilient link between London and New York.
The link is irrelevant. For this level of accuracy you will need GPS
receivers, or atomic clocks, hand carried from a standards lab. (Note
the 90 percentile figure quoted for GPS is about 0.05 microseconds, and
normal receivers tend to claim about 1 microsecond.)
Do you have dark fibre all the way? Are the cables balanced with regard
to thermal expansion in the two directions?
> Searches on the web have made claims that NTP can achieve this accuracy.
Maybe, but you don't get to specify the OS and you won't be allowed to
run anything but NTP on the machine!
> Unfortunately the sales rep for the NTP server we looked at told me that the best I
> could expect is 2-5 ms synchronization across servers.
Reasonable expectations, on Unix, with no GPS receiver.
More information about the questions